This project would develop robust statistical methods for describing intraindividual and interindividual variability in workplace respirator penetration values with two risk assessment/management applications in mind: (1) to establish a sound statistical basis for determining a respirator's assigned protection factor and maximum use concentration; and (2) to deduce the distribution of inhaled contaminant concentrations given data-based distributions of ambient contaminant levels and respirator penetration values. The project has five interwoven components. First, we would identify and compare statistical models appropriate for describing data sets of respirator penetration values. We would focus initially on one-way random effects models of the log and logit transformed penetration values. Second, we would use simulation to examine the distribution of inhaled contaminant concentrations generated by the combined distributions of ambient contaminant concentrations and respirator penetration values. The analysis would assess the effect of intraindividual and interindividual variability in both the ambient contaminant and respirator penetration variables. Third, we would develop a general statistical framework whereby an APF, for a given model of respirator, can be determined based on specifying: (i) the toxicologically appropriate parameter theta of the intraindividual penetration distribution; (ii) an appropriate percentile of the interindividual distribution of parameter theta; and (iii) an upper (1-alpha) percent confidence limit on the estimate of this percentile from sample data. We would compare confidence limits (APF's) based on parametric methods and variance estimates that assume large sample sizes, with those based on a relatively assumption-free bootstrap method. Fourth, we would develop strategies for combining APF estimates from several studies involving the use of similar respirators under similar conditions; our initial strategy would involve computing averages using various weighting schemes and associated confidence intervals. Fifth, we would develop methods based on a compound beta distribution model to describe data sets for which penetration distributions are so skewed that the log and logit transformations do not yield symmetric distributions.